FiNER-ORD (Financial NER)

Figure 1.

Goal

The FiNER-ORD dataset is a high-quality financial Named Entity Recognition (NER) dataset. Its primary goals are:

  • To address the lack of open-source, manually annotated NER datasets in the financial domain.

  • To benchmark NER models on financial news articles.

  • To identify standard entities (PER, LOC, ORG) in a specialized context.

Description

FiNER-ORD consists of financial news stories collected from Webz.io, manually annotated for named entities.

Dataset Composition

  • Source: Financial news articles from Webz.io.

  • Selection: 201 manually annotated articles selected from a larger corpus of 47,851 documents.

  • Entities:
    • PER (Person)

    • ORG (Organization)

    • LOC (Location)

  • Modality: Text

  • Language: English

Annotation Process

The dataset was manually annotated by experts to ensure high quality, serving as a gold-standard benchmark for financial NER.

Example

Below is a representative example of the NER task in FiNER-ORD:

Table 4 FiNER-ORD Examples

ID

Text

Entities

0

Elon Musk, CEO of Tesla, announced the new factory in Berlin.

PER: Elon Musk, ORG: Tesla, LOC: Berlin

1

Apple Inc. shares rose after the report released in Cupertino.

ORG: Apple Inc., LOC: Cupertino

Task Description

Named Entity Recognition

  • Input: A sentence or paragraph from a financial news article.

  • Output: A sequence of tags indicating the start and type of entities.

  • Challenge: Distinguishing between financial entities in complex news narratives.

Evaluation Metrics

  1. F1-Score: The primary metric for evaluating NER performance.

  2. Precision: Accuracy of the predicted entities.

  3. Recall: Ability to find all relevant entities.

Why Use FiNER-ORD?

  • Real-World Data: Based on actual financial news, reflecting real-world usage.

  • High-Quality Annotations: Manually curated to minimize noise.

  • Open Benchmark: Freely available for research, fostering community progress.

References

For dataset access and more details, visit the HuggingFace dataset page.